Add XArray compatibility features

[hameerabbasi]

Once the right broadcasting setup was in place; it was trivial to implement the three-argument version of where and NaN-skipping aggregations.

cc @mrocklin Feedback welcome.
cc @shoyer Feedback really welcome, as you know the ins and outs of XArray. Of course, "I can't" is okay; as always. :-)

[hameerabbasi]

Tests not yet added.

[hameerabbasi]

Unfortunately, there doesn't seem to be a way to override np.nansum. I added a print command in there, and:

>>> import numpy as np
>>> import sparse
>>> x = np.asarray([5, 6, np.nan])
>>> s = sparse.COO.from_numpy(x)
>>> s.nansum()
nanreduce
11.0
>>> np.nansum(s)
11.0

Edit: It doesn't want to fall back to our implementation even if we err when coercing.

>>> class NoncoercibleCOO(sparse.COO):
...     def __array__(self, *args, **kwargs):
...         raise ValueError('Cannot coerce COO.')
...     
>>> s = NoncoercibleCOO.from_numpy(x)
>>> np.nansum(s)
Traceback (most recent call last):
  File "<input>", line 1, in <module>
  File "/Users/hameer/anaconda3/envs/sparse/lib/python3.6/site-packages/numpy/lib/nanfunctions.py", line 581, in nansum
    a, mask = _replace_nan(a, 0)
  File "/Users/hameer/anaconda3/envs/sparse/lib/python3.6/site-packages/numpy/lib/nanfunctions.py", line 64, in _replace_nan
    a = np.array(a, subok=True, copy=True)
  File "<input>", line 3, in __array__
ValueError: Cannot coerce COO.

[hameerabbasi]

Hmm. It seems that (for nanmin and nanmax), sometimes we get -inf as a value (on a full axis that contained JUST nan and nothing else), whereas NumPy actually returns nan. I think the actual value of the reduction in this case is debatable (I, for one, think empty min should be inf, etc, as a hobbyist mathematician).

In any case, I'm willing to leave this as an xfail for now. Any input welcome.

[hameerabbasi]

I added tests and matched the Numpy API. Looks review ready to me. 💃 cc @mrocklin @shoyer

[hameerabbasi]

Do we want to support the edge case of NaNs in object arrays? It's proving difficult. It's certainly possible, but only with a bit of trickery.

[mrocklin]

Make a test with pytest.mark.xfail and move on?

…

On Tue, Feb 20, 2018 at 4:08 PM, Hameer Abbasi ***@***.***> wrote: Do we want to support the edge case of NaNs in object arrays? It's proving difficult. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#102 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AASszHNoWLT6YB86IzQaL6O_9tQ6yww2ks5tWzRGgaJpZM4SKZV8> .

[shoyer]

I don't think there are many uses for object dtype sparse arrays. Mostly we use those for strings, but I imagine sparse arrays are most useful for numbers.

…

On Tue, Feb 20, 2018 at 1:09 PM Hameer Abbasi ***@***.***> wrote: Do we want to support the edge case of NaNs in object arrays? It's proving difficult. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#102 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABKS1lF-wLH1pZ4iTlLns0ZNY24qywqnks5tWzRGgaJpZM4SKZV8> .

[hameerabbasi]

It seems we don't support object arrays in many cases at the moment. I've opened #104 to track this but am inclined to ignore this.

[hameerabbasi]

I fixed the nan issue (by using fmin/fmax); tested for it; and matched the Numpy warning (and tested for it). This time around I feel this really is ready for review.

Edit: I looked at the implementation for nanmin etc. in Numpy. Unfortunately, without hooking into Numpy and replacing its functions with our own if the args are SparseArray, I don't see a way to solve this.

[mrocklin]

This looks good to me :)

[hameerabbasi]

Merged!